The confusion of correlation with causation represents a real issue for management science and is endemic in the practice of knowledge management; indeed the whole issue of causation in social systems is problematic. The issue is well summarized in a metaphor from Christensen & Raynor (“Why Hard-Nosed Executives Should Care About Management Theory” Harvard Business Review, September 2003, pp. 67-74) as follows:                Imagine going to your Doctor because you're not feeling well.        
Before you've had a chance to describe your symptoms, the doctor writes out a prescription and says “take two of these three times and day, and call me in a week.”                “But—I haven't told you what's wrong,” you say. “How do I know this will help me?”        “Why wouldn't it” says the doctor. “It worked for the last two patients.”        
No competent doctor would ever practice medicine like this, nor would any sane patient accept it if they did. Yet professors and consultants routinely prescribe such generic advice, and managers routinely accept such therapy, in the naïve belief that if a particular course of action helped other companies to succeed, it ought to help theirs too.
A complex adaptive system (“CAS”) is a system that is complex (diverse and made up of multiple interconnected elements) and adaptive (has the capacity to change and learn from experience). A CAS has high levels of uncertainty in which outcomes are inherently unknowable. The same thing only happens the same way twice by accident, and the system is highly vulnerable to massive outcome changes based on small perturbations in ways that cannot be predicted. A CAS can be best understood and managed as an evolving system. This requires a rethinking of risk management from probabilistic models based on possible outcomes, to an understanding of the degree of stability and volatility within the system. In addition, cognitive science has transformed our understanding of how humans make decisions, challenging the model of human decision making as one of rational choices based on personal self interests, to one based on matching patterns acquired through experience or through the transfer of narratives and other fragmented material.
Complex Adaptive Systems
Traditionally, systems have been viewed as falling into either of two broad categories:                Ordered systems are systems in which there are repeating relationships between cause and effect that can be discovered by empirical observation, analysis, and other investigatory techniques.                    Once those relationships are discovered, we can use our understanding of them to predict the future behavior of the system and to manipulate it towards a desired end state.                        Chaotic systems are systems in which the agents are unconstrained and present in large numbers. For this reason, we can gain insight into the operation of such systems by the application of statistics, probability distributions, and the like. The number and the independence of the agents allow large number mathematics to come into play.        
A third type of system is a CAS, wherein agent behavior is loosely constrained by the system, but in turn, the agents modify the system through their interactions with themselves and the system. In this context, an agent is defined as anything that acts (it can be an individual, a group, an idea, a regulation, etc.). The mutual interaction produces an inherently unpredictable system. A CAS is highly susceptible to minor changes or weak signals, sometimes illustrated by the cliche of the flapping butterfly's wing. However, given the multi-faceted nature of systems and these broad definitions, systems may have characteristics that stretch across multiple categories of these above-described system types.
As constraints in an ordered system change, a system can shift to being complex, or even collapse into chaos. For example, attempting to exert excessive control through bureaucracy may build up tension through frustration, which can lead to a collapse of control and increased levels of fraud.
A CAS can appear chaotic or, with the benefit of hindsight, as ordered. This latter case is described as retrospective coherence. After something has happened, it is easy to see the significant pattern of events, but detecting the pattern in advance may be virtually impossible using prior art information systems. Detecting relevant patterns early is known as weak signal detection.
Most (if not all) human systems are CAS. We respond and adapt to local interactions and we are constrained by systems, but we are also capable of modifying those systems. Small inputs or perturbations may lead to unintended and unforeseeable consequences. Once sufficiently disturbed, such a system is altered irreversibly and will not return to the previous equilibrium state. If we reduce the system constraints, increase agent interactivity, and increase the variety of those agents and their environment, then radically new patterns become possible. If we understand that a system is a CAS, then our expectations of decisions are different. We are not making decisions based on forecastable outcomes and best practice, as both are impossible. We cannot adopt an approach based on fail-safe design, but have to switch to safe-fail experiments and monitor for the emergence of patterns. Some patterns we amplify and some we dampen depending on the evolutionary direction we wish the system to take. We thus influence the evolution of the system towards an unknowable future state; we do not waste energy in trying to achieve a predefined system outcome.
The Basis of Human Intelligence
Klein (Klein, G., “Sources of Power: How People Make Decisions” MIT 1998) established in his research into decision making that humans make decisions on a first fit pattern match, either with past or hypothecated future experience. The choice of patterns is one of satisfying, not optimizing; it is not the best fit, but the first fit patterns that are used. This is radically different from the information processing, rational decision maker that is typically assumed. We do not scan all available information, but perhaps only 5-10%. Based on this partial scan, we match against patterns stored in our memory and perform a first fit pattern match against those patterns.
In addition, the idea of distributed cognition is central to the application of complexity to human systems and has profound implications for knowledge management. Complexity based approaches to management handle ambiguity and uncertainty by avoiding central control and allowing high levels of agent interaction to create emergent patterns of meaning.
Humans as Fragmented Processors
Humans are pattern processors. Our response to experiences, in particular those of tolerated failure, create vivid patterns through which we filter data and make decisions. A major distinguishing feature of human intelligence is our propensity to create cultures that increase familial and tribal bonds, and to pass on knowledge other than through genetic evolution and experience: we are, at our very essence, storytellers. The greater part of our evolutionary history has been spent in an oral tradition, and the modern environment of social computing, comprising multiple fragmented conversations, can be viewed as a return to (or arguably just a continuation of) an oral tradition.
Stories and other fragmented material are also fractal in nature and are linked to common work and social group experience. When a family assembles for a wedding or funeral, the family members will retell the identity stories of their family. The same is true of work groups, organizations, and cultures—all of which are self-similar and provide a capacity for common action. Engineers working on a long-term project create stories that define the experience and key learnings that they derive. Mentors tell stories of their own experiences to those they mentor and those mentored, in their turn, modify those teaching stories and create their own. To understand what we know and how we know it, and by implication how we make decisions, we need to understand the multi-facetted narratives of our day-to-day discourse.
A broader definition of fragmented material (“fragments”), also called “information objects”, includes anything that allows people to make sense of the world: paintings, pictures, sacred objects, blogs and the like. Naturally occurring stories typically come as fragmented anecdotes. Occasionally you get a fully formed and developed story, but mostly they are anecdotal, often only one or two paragraphs long when transcribed. Paintings and pictures are often a better form of fragmented expression than a pure story in textual form. A story is always told in a context, from a context. It will trigger a reaction that is not necessarily consistent with what was intended by the storyteller. Each reader has his or her own context and situation.
Semantic Approach
A consistent belief among many in the western world is that everything is reducible to its machine-like components. Semantic approaches, as exemplified by visions of the Semantic Web—an evolving extension of the World Wide Web, use highly ordered building blocks and elaborate algorithms. Semantic approaches are typically directed to standardization, formats and microformats, and strive to be explicit and unambiguous in characterizing objects and their relationships to other objects, through tags, ontologies, Resource Description Framework, taxonomies, and the like.
Natural language processing is an additional semantic approach where software attempts to understand the meaning of a piece of text. The fundamental ambiguity of language, its dependence on context for meaning, and other complexities (e.g., sarcasm, poor spelling, poor grammar, dialects, and unconventional writing styles), make it impossible to automate the process of extracting anything beyond superficial meaning from an individual piece of text.
Boisot (Boisot, M., Knowledge Assets Oxford 1998) argues that the most valuable knowledge is codified (to be easily shared and re-used), undiffused (proprietary) and abstracted (key learnings are taken from the world and useful understandings and models are created that can be used in other contexts to make sense of a new environment and to aid in decision making). The least valuable knowledge is un-codified (difficult to share and re-use), diffused (widely-distributed, non-proprietary), and concrete (this can be thought of as very specific descriptions with no abstraction of meaning, similar to raw field intelligence). Using this model, the current internet is widely diffused and has a reasonable level of codification, but generally lacks abstraction. The Semantic Web will significantly increase the level of codification but will only slightly increase the level of abstraction—hence, only marginal net improvement in net knowledge value.
There remains a need, therefore, for increasing the abstraction (allowing re-contextualization or blending codified models to different contexts) of information objects, with improvements to codification for creating more valuable knowledge assets.
Early related work by one of the present inventors and others is described in U.S. Pat. No. 7,136,791, which is hereby incorporated by reference. In particular, specific methods for eliciting narrative materials are disclosed therein.
U.S. Patent Publication No. 2004/0006567, for which one of the present inventors is also a co-inventor, is also incorporated by reference.